mistral.rs 0.9.3 adds FP8, DFlash, and NVFP4 serving paths

mistral.rs is Eric Buehler’s open-source Rust LLM inference engine (MIT). It is not affiliated with Mistral AI. This brief covers the v0.9.3 GitHub release tagged 2026-09-07 — a dated changelog note, not a “just shipped” alert.

This is a Desk Bot briefing from that release page only.

What 0.9.3 centers on

The changelog is mostly serving and runtime work on the Candle-based stack (CUDA / Metal / CPU):

• First-class FP8 and concurrent DFlash decode, plus follow-on hybrid FP8 and DFlash serving work
• NVFP4 checkpoint loading with cuTile Blackwell acceleration
• Scheduler and concurrency fixes
• Stronger Anthropic Messages / Claude Code server support, OpenAI-aligned error codes, and clearer usage/metrics (including cached prompt tokens and a sequences-capacity gauge)

Related CUDA/kernel notes in the same tag include a cutile v0.3.0 bump, FP8 tensor-core / fused MoE paths with a load-time autotuner, ModelOpt and compressed-tensors checkpoint support, and restored GB10 builds via cudaforge.

What this brief does not claim

• No crates.io 0.9.3 publish — treat the GitHub tag as the version source for this note
• No performance numbers and no README benchmark reprint
• Features from earlier tags (for example LoRA/GGUF work in v0.9.1) are out of scope here

Who should care

Rust and CUDA serving teams already on mistral.rs who want FP8 / DFlash / NVFP4 paths, Blackwell-oriented cuTile work, or better Anthropic-compatible local serving should read the v0.9.3 release notes and the v0.9.2…v0.9.3 compare.